ShellFrame v0.40.0

v0.40.0 (2026-10-18)

Added

• Cursor Agent is a supported tab type, beside Claude Code, Codex, and Grok Build.
  A cursor-agent tab was only a shell command: no status dot, no model badge,
  no way to reopen the same chat after a reboot, and the phone footer could not
  tell a running turn from an idle prompt. The installer also creates an agent
  binary that is Grok Build, and the desktop IDE is a different cursor command,
  so matching either of those would attach the wrong provider. ShellFrame now
  recognises only the cursor-agent program, including its Windows launcher.
  A new tab is started with a pinned session id and reopened with resume when
  that chat still exists on disk. The status dot comes from the chat transcript
  and from the on-screen line that says how to stop the turn; the confirm-to-exit
  line is idle, not busy. The usage view names the model and whether it sits
  inside the subscription, spends prepaid credit as a third-party model, or is
  left on automatic selection. There is no local quota file, so the pill does
  not invent a percentage. The new-tab preset pins a subscription model and does
  not enable on-demand billing. Covered by tests_cursor_provider.py.

  Cursor Agent 成為支援的分頁類型，與 Claude Code、Codex、Grok Build 並列。
  之前這類分頁只是一般指令：沒有狀態燈、沒有模型徽章、重開機後不能接回同一段
  對話，手機上的忙碌列也分不出回合還在跑。安裝器另外建立的 agent 指令其實是
  Grok Build，桌面 IDE 則是另一個 cursor 指令，認錯就會套上錯誤的 provider。
  現在只認 cursor-agent 這個程式，包含 Windows 的啟動器。新分頁會指定 session
  id，記錄還在時用 resume 接回。狀態燈來自對話記錄，以及畫面上那行怎麼停止回合
  的字；再按一次才離開的確認行算閒置，不算忙碌。用量畫面會寫出模型，並標它是
  訂閱內、會花預付額度的第三方模型，或是自動選模型。本機沒有配額檔，所以用量
  膠囊不編一個百分比。新分頁的預設釘在訂閱模型，也不打開按需計費。由
  tests_cursor_provider.py 守著。

• A tab can hand work to an agent on this machine or on a paired machine, and get one result back.
  Agents could already delegate to a local role, but a task that belonged on
  another paired computer had no command that both delivered the prompt and
  expected a report. The new commands send a message, hand off a task, return
  a result, and print the audit log. Delivery reuses Frame Link. The peer must
  already be paired, an unknown name is refused, and a one-way pairing still
  cannot inject a prompt into the controlling machine. Nothing is stored for
  later when the peer is offline, so a message cannot arrive twice after the
  link returns. The same text inside one minute, and the same id, are dropped.
  A result is delivered once. A handoff that has already reached the depth cap
  can still report. The feature stays behind the existing experimental flag,
  which remains off until it is turned on. Covered by tests_agent_link.py and
  tests_frame_link.py.

  分頁可以把工作交給這台或已配對機器上的 agent，並收回一次結果。
  先前只能把工作派給本機角色。該由另一台已配對電腦做的事，沒有一個指令會把
  提示送過去並等對方回報。新指令可以傳訊息、移交任務、回傳結果、印出稽核紀錄。
  傳送走既有的 Frame Link。對方必須已經配對，不認識的名字直接拒絕，單向配對
  仍然不能把輸入送進主控那台。對方不在線就不先排進待送佇列，避免連線恢復後送出
  兩次。同一分鐘內的相同內容，以及同一個編號，都會丟掉。結果只送達一次。深度
  已經到頂的移交仍可以回報。功能留在原本的實驗旗標後面，預設維持關閉，要打開
  才會送。由 tests_agent_link.py 與 tests_frame_link.py 守著。

• Delegate and handoff now carry a usage hint, and a separate command prints that hint without sending anything.
  Choosing an agent by habit could spend a nearly exhausted weekly allowance,
  or pick a third-party model that draws prepaid credit, while a cheaper local
  or subscription agent would have done the job. The hint classifies the task
  as light work, coding, or planning, then suggests the cheapest adequate
  option from the readings already on the usage pill. Unknown usage stays
  eligible and is not treated as zero. A planning model whose week is at least
  nine tenths used is left out. Other agents at that level move to the back of
  the list. If the task names a third-party model, the hint says that choice is
  paid and does not select it. The suggestion is attached to the reply. It does
  not change the destination and it does not start a process. Covered by
  tests_agent_route.py.

  派工與移交現在會附上用量建議，另有一個指令只印同樣的建議、什麼都不送。
  依習慣挑 agent，可能把快用完的週配額再花掉，或選到會扣預付額度的第三方模型，
  而更便宜的地端或訂閱 agent 就做得完。建議先把任務分成輕量、寫程式或規劃，
  再用用量膠囊上已有的讀數，挑夠用又較省的一個。讀數未知仍可被選到，不會被當成
  已經用了零。規劃用的模型若這一週已用到九成以上就排除。其他 agent 到了這個
  水位只是排到後面。任務若點名第三方模型，建議會註明那是付費，並且不選它。
  建議附在回覆上，不改目的地，也不會因此啟動行程。由 tests_agent_route.py 守著。